Prof. Kalin Kopanov

Bulgarian Academy of Sciences, Bulgaria



Title of the paper: Scaling Quantum Support Vector Machines for AI-Generated Text Attribution: Dimensionality, Shots, and Linguistic Feature Insights


Abstract: Attributing AI-generated text to its source model remains challenging. We evaluate Quantum Support Vector Machines (QSVMs) using a controlled setup that standardizes preprocessing and SVM parameters to ensure fair, reproducible comparisons with a calibrated linear baseline. On a balanced dataset of 5,800 texts, the baseline employs 3,018 TF-IDF and linguistic features, while the quantum models work with PCA-reduced inputs of 4-10 dimensions. We test two simulators: a tensor-network backend with sampling (cuTensorNet) and a deterministic state-vector backend. Our evaluation uses a fixed 20% hold-out set, and we report ROC/PR-AUC, Brier score, and calibration on a separate 200-sample test set. The linear model achieves approximately 98% accuracy. Quantum models reach around 60–66% accuracy on cuTensorNet, with training and hold-out data closely aligned, whereas the deterministic simulator tends to overfit. Dimension plays a bigger role than shot count (512-4,096). Moving from 4D to 10D improves results; 8D provides the clearest separation, and increasing shots beyond 1,024 adds minimal benefit. Across all settings, the quantum models consistently enhance stylistic and sentence-structure cues such as capitalization, punctuation balance, and verb or question usage, supplementing the baseline’s stronger lexical signals.

Bio: To be announced soon




Categories